ChatGPT Images 2.5: precise edits matter more than a prettier first image
October 8, 2026 update: Nano Banana 2.1 and image pipeline migration
Marek Bartoš’s October 7 video highlighted Nano Banana 2.1. The original release predates the video: the English Gemini API changelog records general availability on October 6, 2026, alongside deprecation of gemini-3.1-flash-image.
The shutdown date needs care. When checked on October 8, the English changelog gave no date, while the Spanish version stated October 29. I would not treat that conflicting date as a confirmed deadline; check Google’s current guidance before migrating.
The model documentation lists gemini-nano-banana-2.1, 1K, 2K and 4K output, and Batch API support. Google claims better text rendering and consistency across repeated edits. These are vendor claims, not results from my own comparison.
I would add the model to the same marketing pipeline trial as Images 2.5 and Ideogram. Choose ten approved product photos. Apply three successive edits to each: the background, one detail’s color and a short Czech caption. After every step, check product geometry, the logo, diacritics and unwanted changes outside the requested area. Include a wide banner to expose edge artifacts.
In n8n or Make, separate image API calls from publishing. Keep the model and parameters in configuration, and record the source asset ID and prompt version for each run. Switch a copy of the workflow first, saving outputs to a separate folder for human approval. Choose the default model only after measuring cost per approved asset, error rate and correction time. A new release justifies a migration test; it does not establish a winner.
October 3, 2026 update: testing Ideogram 4.5 for targeted repairs
Zaujaloma AI’s October 2 roundup highlighted Ideogram 4.5. Its product page claims reduced drift across successive edits. The comparison gallery is vendor evidence, not my benchmark.
The Precise Edit endpoint is relevant to an existing pipeline: its documentation describes original-size output and mask support. The API overview also covers asynchronous processing and warns that result URLs expire.
This reinforces the article’s argument: choose a model for the repair you need. It does not justify moving an entire Images 2.5 workflow automatically.
I would run a small comparison using the same approved product photographs. Each branch would perform the same sequence: replace the background, recolor one component, and change a short text label. After every step, compare the logo, product geometry, and untouched areas against the original. Check Czech diacritics and legibility at the final display size too.
In n8n or Make, keep the original, mask, instructions, and every output version under one job ID. Copy completed assets into your own storage; a generator’s temporary URL should never become the CMS asset URL. Evaluate cost per approved visual, including retries and human review. I have not run this pilot, so I am not declaring a winner or claiming pixel fidelity on my own data.
OpenAI released ChatGPT Images 2.5 on September 8, 2026. The visible improvements are sharper detail, more natural lighting, richer texture, and faster generation. The more important business change is preservation: when you edit one part of an image, the model is more likely to keep the person, product, composition, and brand treatment intact.
That is the line between an impressive image generator and a dependable creative workflow. Models could already produce a strong first image. The expensive part was the second, third, and sixth revision, when changing a background also damaged a face, product package, or approved layout.
What version 2.5 changes
OpenAI reports up to 50% lower generation latency than Images 2.0. The model is designed to make more focused edits and preserve earlier decisions over a longer multi-turn session without progressively degrading the asset. Transparent backgrounds, complex layouts, and prompts involving real-world information have also improved.
ChatGPT adds four practical controls:
- Sketch turns a rough drawing made inside ChatGPT into a visual reference,
- templates provide a starting structure for formats such as posters, merchandise, and product photos,
- comments can be placed directly on an image to target a specific edit,
- shared images can include their prompt so a colleague can reuse the approach with different source material.
Images 2.5 is rolling out across ChatGPT, ChatGPT Work, and Codex on desktop, mobile, and web. Teams building governed pipelines should pay closer attention to the two new API models.
Flare for volume, Sunburst for expensive exceptions
GPT-Image-2.5 Flare is OpenAI's default recommendation for most applications. It is positioned as higher quality than GPT-Image-2 at half the latency. That makes it the sensible worker for product variants, social content, visual search, transparent cut-outs, prototypes, and higher-volume generation.
GPT-Image-2.5 Sunburst is the slower specialist tier for work where precision and edit stability matter more: production campaign creative, polished product imagery, or a sensitive revision of an approved asset.
The current API rate card lists the same per-million-token prices for both models:
| Modality | Input | Cached input | Output |
|---|---|---|---|
| Image | $8 | $2 | $30 |
| Text | $5 | $1.25 | — |
For comparison, the older GPT-Image-2 is listed at half those rates: $4 image input, $1 cached image input, $15 image output, and $2.50 text input. A token rate is not the price of one image. Actual cost depends on input size, resolution, quality, number of variants, and repeated edits.
The useful metric is therefore cost per approved asset. Flare can still make the overall workflow cheaper if better preservation and lower latency reduce rejected variants and manual rework.
A practical production workflow
For a marketing or ecommerce pipeline, I would separate five stages:
- Store the brief as structured data: purpose, audience, product, required brand elements, prohibited motifs, format, and destination channel.
- Use Flare to create several lower-cost directions and composition variants.
- Automatically validate dimensions, transparency, required product presence, and safety rules; have a person verify visual truthfulness and brand compliance.
- Escalate only selected assets to Sunburst or manual finishing.
- Record the prompt, source references, model, parameters, cost, comments, and final approval.
In n8n or Make, a new product record can trigger the first generation, save variants to a DAM or Drive, and create an approval task. The automation should not publish images involving people, health claims, prices, or legally significant copy without review. The approval gate is part of the architecture, not paperwork added at the end.
Precise editing changes creative economics
The most useful outcome of Images 2.5 will not be another impressive gallery. It will be changing a background, product, piece of copy, or format without breaking the rest of the approved design. That shortens the path from an initial concept to a consistent asset set.
The original conclusion still holds: AI is a strong prototyping and variation layer, not an uncontrolled final-asset factory. OpenAI continues to use C2PA metadata and invisible watermarking, but provenance does not resolve source licensing, consent from depicted people, product accuracy, or brand review.
Sources: OpenAI: ChatGPT Images 2.5, OpenAI API pricing, Images 2.5 system card, the original ChatGPT Images 2.0, and Czech video signals about the previous release from Zaujaloma AI, Tomáš AI, and Marek Bartoš.